A technical paper titled “Traits in Vitality Estimates for Computing in AI/Machine Studying Accelerators, Supercomputers, and Compute-Intensive Purposes” was printed by researchers at SLAC/Stanford College and MIT.

Summary:
“We study the computational vitality necessities of various programs pushed by the geometrical scaling regulation, and rising use of Synthetic Intelligence or Machine Studying (AI-ML) over the past decade. With extra scientific and expertise purposes primarily based on data-driven discovery, machine studying strategies, particularly deep neural networks, have turn out to be extensively used. As a way to allow such purposes, each {hardware} accelerators and superior AI-ML strategies have led to the introduction of recent architectures, system designs, algorithms, and software program. Our evaluation of vitality traits signifies three essential observations: 1) Vitality effectivity as a result of geometrical scaling is slowing down; 2) The vitality effectivity on the bit-level doesn’t translate into effectivity on the instruction-level, or on the system-level for quite a lot of programs, particularly for large-scale AI-ML accelerators or supercomputers; 3) On the utility stage, general-purpose AI-ML strategies will be computationally vitality intensive, off-setting the features in vitality from geometrical scaling and particular goal accelerators. Additional, our evaluation offers particular pointers for integrating vitality effectivity with efficiency evaluation for enabling high-performance and sustainable computing sooner or later.”

Discover the technical paper here. Revealed October 2022. Submitted to Proceedings of IEEE Convention on Excessive Efficiency Excessive Computing (HPEC) 2022.

arXiv:2210.17331v1, Sadasivan Shankar, Albert Reuther.

Associated Studying:
AI Power Consumption Exploding
Exponential enhance just isn’t sustainable. However the place is all of it going?
11 Ways To Reduce AI Energy Consumption
Pushing AI to the sting requires new architectures, instruments, and approaches.
MIT: Computing Power Impact On 5 Domains
How To Optimize A Processor
There are at the least three architectural layers to processor design, every of which performs a big position.



Source link

LEAVE A REPLY

Please enter your comment!
Please enter your name here